What Should a Small Shop Automate, and in Which Order?

Retail businesses in the United States report using AI at 14%, against a national average of 19.8%. Shops are behind, and the reason is not reluctance. It is that most advice about automating a small business is written by people who have never had to pack an order between two support calls.

How many businesses like yours are actually doing this?

Fewer than the noise suggests, and the two best datasets agree. The US Census Bureau's Business Trends and Outlook Survey put national AI use at 19.8% as of 3 May 2026, with firms of 250 or more employees at 37% and firms under 20 employees below 20%. Retail trade reported 14% current use and 17% expected within six months. Information sat at 39.7%.

Across the Atlantic the picture is close enough to be reassuring. Eurostat recorded 19.95% of EU enterprises using AI technologies in 2025, breaking down to 17% of small enterprises, 30.36% of medium ones and 55.03% of large ones. Two statistical agencies, two methods, two continents, one answer: roughly one business in five, and the smaller you are the less likely it is to be you.

The gap between the smallest and the largest is the interesting column. Big firms are not ahead because they have better models. They have people whose job is to set this up. In a shop run by one or two people, that person is you, on a Sunday, and the scarce resource is attention rather than technology. Which is exactly why the order of operations matters more than the tool.

What is the difference between automating and assisting?

One finishes the job. The other hands you a draft. Confusing them is the single most expensive mistake in this whole area, because they carry completely different risks and completely different savings.

Anthropic's own usage research gives a sense of where real use sits. In its June 2026 Economic Index report, the most common outputs from consumer conversations were explanations at 17%, documents and reports at 15% and guidance at 11%. Those are drafts and answers, not actions. The same research found a much higher automation share in API traffic than in chat, which is a polite way of saying that when a machine acts alone, it is usually because a developer wired it to.

For a shop, the practical translation is a rule. Anything that touches a customer, money or stock starts as assist. Anything that produces an internal artefact can start as automate. You move a task from assist to automate only after it has stopped needing edits, and you find that out by counting rather than by feeling.

Which task goes first?

Score each candidate on three things and the argument ends. How often does it happen. How similar is each instance to the last one. What does it cost when it is wrong. High frequency, high sameness, low cost of error goes first. Low frequency, high variability, high cost of error never goes at all.

Read the top row and the bottom row together. The top one saves you twenty minutes a day and cannot really hurt you. The bottom one saves you an hour a quarter and can hurt you badly. Businesses reliably start at the bottom, because pricing feels strategic and delivery questions feel beneath them. The order in that table is deliberately the opposite of the order most people choose.

What about the tasks that look automatable and are not?

Three keep coming up. Bookkeeping reconciliation looks repetitive and is full of one off judgements about what a payment actually was. Supplier negotiation looks like correspondence and is a relationship. And anything where the input is a photograph of something physical, a damaged parcel, a fabric swatch, a part number stamped on metal, will work impressively in a demo and then fail on the fifth real example in bad light.

The tell is variability in the input rather than complexity in the task. A machine handles a complicated but consistent job better than a simple but inconsistent one, which is the reverse of how humans find work hard, and it is why intuition misleads here.

Does this work for a service business rather than a shop?

The scoring survives intact, and the candidate list changes shape. A service business generates fewer product records and far more correspondence, which moves the centre of gravity toward scheduling, quoting and follow up.

The highest scoring task in most one person service businesses is the same one: turning an enquiry into a structured quote. It happens often, the shape barely varies, and a wrong draft costs nothing because you read it before it goes. The lowest scoring is deciding whether to take a job at all, which happens rarely, varies completely and can cost you a month of your life if it goes wrong.

The one that surprises people is appointment reminders. They feel too trivial to bother with, they are the highest frequency task in the whole business, and the effect of getting them right shows up directly in the number of people who turn up.

What does the checking cost?

More than anyone budgets, and it never goes to zero. Automating a task does not remove the work, it moves it from producing to checking, a shift we went through in detail in the piece on how AI moved the work from producing to checking. Producing a product description takes fifteen minutes. Checking a generated one takes four. That is a real saving of eleven minutes, not fifteen, and if the generated version needs a rewrite one time in five, the real saving is smaller again, which is the arithmetic behind working out whether an AI tool actually saved you money.

So measure two numbers per task from the first day. Time per instance after automation, including the check. And correction rate, meaning how often you had to change something before it went out. A task whose correction rate is falling is a task you can eventually release. A task whose correction rate is flat after a month is a task that should go back to assist, permanently, without embarrassment.

Where does the time actually go in a shop?

Nobody can answer that from memory, which is why the first week of this project involves no AI at all. Keep a note, on paper if that is easier, of every task you repeat. Not a time and motion study. Just a tally with a rough duration.

What that week produces is the only input the scoring table needs, and it usually surprises the owner. The tasks that feel heaviest are often rare. The tasks that consume the week are small, frequent and invisible, which is precisely the profile that automates well. A shop that skips this step ends up automating whatever it was annoyed about most recently, which is a different thing from whatever costs it most.

If the output of that week is a set of steps you keep repeating, you have also accidentally written your operating procedures, which is the asset described in getting the business out of your head and onto one page. Those two projects are the same project, done once.

Does the tooling matter as much as people think?

No, and the spending pattern proves it. Most of what a small shop needs is already inside software it pays for, and the surrounding costs are covered in what a small business really spends on AI each month. A separate purchase is justified when a task is frequent enough that the manual version is measurably eating your week, and not before.

Two exceptions where a dedicated tool earns its keep quickly. Document extraction, because pulling structured data out of supplier paperwork is a genuinely hard problem that general chat handles badly, and we walked through the shape of that job in getting line items out of a thousand supplier PDFs. And customer support routing, where the value is in triage rather than in the answer, and the starting point is described in deciding which tickets to hand over first.

What should never be automated, whatever it scores?

Four things, and they are worth writing on the same page as your task list, because the scoring table will happily rank some of them highly and it would be wrong.

Anything that states a policy. Refund windows, delivery guarantees, warranty terms, price matching. A generated sentence about your returns policy is a promise you did not make and may still be held to, which is the ground covered in what happens when a chatbot makes a promise.

Anything that says no to a person. Refusing a refund, declining a job, closing an account. These are rare, they are remembered, and they are the moments where a human sentence is worth more than an efficient one.

Anything that must disclose itself and does not. Several jurisdictions now require you to tell people when they are dealing with a machine rather than a person, and the rules are not uniform, which we sorted through in the piece on whether you have to tell customers they are talking to a bot. Automating a conversation without handling the disclosure is a compliance problem wearing an efficiency costume.

And anything where you cannot describe what a correct output looks like. If you cannot write down the standard, you cannot check the work, and an unchecked automated task is not a saving. It is an unmeasured risk with a time stamp on it.

How do you know it worked, six months later?

Three numbers, and none of them is a productivity percentage from a vendor deck.

1. Hours back per week, measured once a quarter. Repeat the tally week you did at the start. If the same tasks now take less time, the number is real. If it feels faster but the tally says otherwise, believe the tally.
2. Correction rate per task. Falling is good. Flat is a signal to move the task back to assist. Rising usually means your inputs drifted, not that the tool got worse.
3. Number of tasks you abandoned. This is the health check nobody keeps. A business that automated four things and abandoned none was probably not honest about the results.

Notice what is absent. Revenue is not on that list, because attributing a sales change to an automated product description is beyond what a small business can measure, and pretending otherwise produces confident nonsense. Time and correction rate you can measure with a notebook. Start there and stay there.

What does a realistic first quarter look like?

Three tasks, not thirty. Week one is the tally. Week two you pick the single highest scoring task and set it up as assist, meaning it drafts and you approve every instance. Weeks three and four you record the two numbers, time per instance and correction rate, and you resist adding anything.

Month two you add a second task and, if the first one's correction rate has fallen below roughly one in ten, you allow it to run without per instance approval on internal work only. Month three you add a third and you review all three together, which is when you will usually discover that one of them was never worth automating and can be dropped.

The output at the end of the quarter is modest and durable: two or three tasks that reliably take less time than they did, a written note of what you tried and abandoned, and a habit of measuring. That is a worse story than the ones you read and a better outcome than most of them produce.

What happens to all this when you hire someone?

It gets more valuable, and in a way that is not obvious until it happens. An automated task comes with a written description of what good output looks like, because you had to write that description to check the work. That description is a training document you did not know you were writing.

The Census figures make the point from the other direction. Firms with 250 or more employees report AI use at 37% while firms under 20 sit below 20%, and the usual explanation is budget. Budget is part of it. The larger part is that a company with staff has someone who can write the standard down, and a company without staff has an owner who holds the standard in their head. Automating a task forces the standard onto paper, which is the same work as preparing to delegate it.

So the sequence for a growing shop runs in an order most people reverse. Automate the boring frequent task. Write the check standard because you need it. Hand both to the new person on day one. What you have handed over is not a chore, it is a chore with a quality bar attached, and new hires get to useful work faster with one than without.

Does automating make the business harder to sell or hand over?

The opposite, if you keep records. A business whose processes live in one person's habits is worth less than one whose processes are written down and running, and that gap is well understood by anyone who has bought a small company. Every task you automate with a written standard moves a little value from the founder's head into the business itself.

The failure case is worth naming though. A shop that automates using a tangle of personal accounts, undocumented connections and a tool nobody else can log into has made the business more fragile rather than less. Keep the accounts in the business name, keep a one page note of what runs where, and keep the list of what you abandoned and why. That note takes ten minutes a quarter and it is the difference between an asset and a liability with good uptime.

Is any of this different if you sell online rather than in person?

The scoring is identical. The candidate list is not, because an online shop generates more structured text than a physical one: product data, order confirmations, listing copy, returns correspondence, feed exports. Structured, repetitive text is the best raw material there is, which is why the retail figure at 14% reads as an opportunity rather than a verdict.

There is also a compounding effect worth naming. Every task you automate well produces a cleaner record of how your business actually runs, and that record makes the next task easier to automate. Shops that build the catalogue and the copy on a consistent structure from the start get this for free, which is part of what our AI store builder is for, and shops that accumulated three years of inconsistent product data pay for it later in cleanup.

One last thing to hold on to. The census figures say four businesses in five are not doing any of this yet. That is not a warning that you are behind. It is a reminder that nobody has solved it, that the advice available is mostly written for companies with staff to spare, and that a shop which automates two boring tasks properly this quarter will be ahead of almost everyone selling the same thing.